FAST Copper For Broadband Access: An Overview

Transcription

1 FAST Copper For Broadband Access: An Overview Mung Chiang Electrical Engineering Department, Princeton With J. Huang, D. Xu, Y. Yi, C. W. Tan, R. Cendrillon Allerton, SPIE, INFORMS September November, 2006

2 What s FAST Copper? >10X improvement in copper-last-mile broadband access through fiber/dsl deployment, engineering innovations, and fundamental research R3Q: Rate (at application level), reach, reliability, quality NSF ITR sponsorship Princeton, Stanford, Fraser Research Lab PI: M. Chiang, Co-PIs: J. Cioffi and A. Fraser Main industry collaborator: AT&T Timeline: : initial work with SBC : formal duration of the project 2008-: continued research and industry adoption

3 Outline Is it possible to get truly broadband with phone line? Architectural issues Frequency Amplitude Space Time FAST and FAST: FAST Copper is different from TCP FAST Research talk: Not focusing on stories about industry deployment Midway report: FAST Copper is just starting to gain full momentum Partial report: Only Princeton s part summarized here

4 Introduction

5 Why Fiber/Copper? Alternatives of broadband access: Wireless: reliability, coverage, and backhaul issues Cable modem: not ubiquitous, bandwidth sharing issues Fiber to the closet: per-customer labor cost prohibitive (especially for brown-field suburban in US) Existing DSL: 160 million users, but not fast enough Fiber/Copper: Best of ubiquity, broadband, reliability, and migration Broadband over fiber and phone wires Example: AT&T s Lightspeed Project

6 Where Are Bottlenecks and Where To Improve Attenuation: Solution from Space Crosstalk: Solutions from Frequency, Amplitude, Time Realistic estimates on improvements coming from research: Frequency: 2X (even more through signal processing) Amplitude: >2X Space: enabler of rate, reach, reliability Time: 2X Not even bringing in wider bandwidth, multiple twisted-pairs, and systems debugging yet

7 Key Ideas It s not a dedicated line, it s a (multi-carrier) interference channel Turn competition to cooperation in frequency and time From low frequency mentality to high frequency mentality It s not a voice line, it s a bursty data and video line Squeeze in more than you have bandwidth for From deterministic mentality to statistical mentality How to make the engineering work? A lot of research (and deployment) challenges

8 Challenges and Connections Two types of challenges Many challenging problems in terms of resource allocation: Information theory: multi-carrier interference channel Signal processing: multi-user transmissions Stochastic theory: statistical multiplexing Graph theory: survivable tree design Optimization theory: nonconvex and globally coupled optimization Networking: resource allocation and Layering As Optimization Decomposition But the biggest challenge is architecture design for broadband access

9 Typical Deployment: Access Part DMT (Discrete Multi Tone) Transmissions IP and PSTN Network TX CO Copper Line RX Customer 1 crosstalk Fiber TX RT RX Customer 2 Downstream Transmission

10 Typical Deployment: End-to-end IO CO VHO VHO IO CO CO VHO VHO SAI VHO SAI SAI IO IO 10 Gbps 1 Gbps CO CO CO SAI SAI SAI 100 Mbps

11 Architectural First Architecture: functionality allocation More influential, harder to change, less understood than resource allocation Metrics: Performance, X-ities, Cost and complexity Modularization: vertical decomposition by a protocol stack Distribution of control: horizontal decomposition into network elements Coupling between horizontal and vertical decompositions Example: who takes care of traffic shaping? Example: Where to do error control: FEC, ARQ, R-UDP, TCP, or application layer?

12 Horizontal Decomposition Video server placement: Tradeoff between response time and scalability Distribution server and cache placement: Where to take care of channel changes? Where are the boundaries of multicast group? Even bigger issue: How big should the access network be? Tradeoff among reliability of access tree, feasibility of big switches, complexity of backbone network, ease of management

13 Vertical Decomposition and Time-Scales of F A S T upper layer traffic shaping Scheduling Time Amplitude Spectrum Mgment Frequency Topology Design lower layer Space Packet Flow Montly/Yearly shorter time-scale longer time-scale Time-scale: Time > Frequency >> Amplitude >> Space Low-complexity Spectrum Management Algorithm: Time Frequency Time-scale separation lowers price of modularity

14 Vertical Decomposition and Time-Scales of F A S T Extreme cases: spatial division multiplexing (S), time division multiplexing (T), frequency division multiplexing (F), turn away users (A) can all tackle crosstalk Possible rate regions attainable (Frequency): determined by deployment topology (Space) Feasibility and stability of scheduling (Time): determined by placement of traffic shapers and schedulers (Space) Two obviously coupled degrees of freedom: Time and Frequency Furthermore, capability of Time: determined by time-scale of Frequency Amplitude control depends on rate region attainable (Frequency) Interesting interaction between Time and Amplitude: next slide

15 Modularity-Performance Tradeoff user 2 System Capacity Region Admission Region of π 2 Admission Region of π 1 faster scheduling time-scale higher computational complexity user 1 A(π 1 ) A(π 2 ), where A(π): admission region of scheduling π Scheduling Algorithm: π 1 and π 2 π 1 : at flow-level time-scale π 2 : at packet-level time-scale exploiting opportunism Conservative admission control A(π 1 ) removes the need for π 2 scheduling

16 Mid-point in the Talk Move from the quantification of architectural tradeoffs to A very brief summary of current progress on F, A, S, T

17 Frequency

18 Dynamic Spectrum Management Question: How to allocate power (bit loading) across different tones and competing users to turn competition to cooperation? Problem formulation: X maximize w n R n {p n 0} n n subject to X k p k n P n, n User n s achievable rate R n = P k log 1 + p k n P m n αk n,m pk m +σk n Total power constraint: P n = p k n 0, k, P k pk n P max n Characterize Pareto boundary of rate region [Centrillon et. al. 04] «Challenging optimization problem: Nonconvex and coupled (across users and across tones)

19 History of DSM algorithms IW: Iterative Water-filling [Yu Ginis Cioffi 02] OSB: Optimal Spectrum Balancing [Cendrillon et. al. 04] ISB: Iterative Spectrum Balancing [Liu Yu 05] [Cendrillon Moonen 05] ASB: Autonomous Spectrum Balancing [Huang Cendrillon Chiang Moonen 06] Algorithm Operation Complexity Performance IW Autonomous O (KN) Suboptimal OSB Centralized O `Ke N Optimal ISB Centralized O `KN 2 Near Optimal ASB Autonomous O (KN) Near Optimal K: number of carriers N: number of users

20 Reference Line Concept Dynamic pricing for dynamic coupling: decouple tones Static pricing for static coupling: decouple users CO CP CO Reference Line CP RT Actual Line CP RT CP RT CP

21 Key Idea of ASB User n solves the following problem: maximize p n 0 w n R n + R ref n subject to X k p k n P n where the reference line rate is: R ref n = X k log 1 + p k,ref α k,ref n p k n + σk,ref! Parameters in red are constants known a priori through channel measurement Autonomous: Only local information is needed Low complexity and achieve near optimal performance

22 ASB Algorithm: Basic Sketch repeat for each user n = 1,..., N end repeat until convergence for each carrier k = 1,..., K, find Find p k n by solving one subproblem for tone k λ n = ˆλ n + ε λ `Pk pk n Pn max + h P w n = w n ε w k Rk n Rn target i + until convergence

23 Typical Result from Realistic Simulator Almost identical to optimal benchmark by centralized computation More than double the rate for typical deployment scenario User 1 User 2 User 3 User 4 CO 2Km RT 3Km 5Km 4Km 4Km RT CP 3.5Km RT CP CP 3Km CP User 1 achievable rate (Mbps) Optimal Spectrum Balancing Iterative Spectrum Balancing 0.6 Autonomous Spectrum Balancing Iterative Waterfilling User 4 achievable rate (Mbps)

24 Typical Spectrum User 2: downstream transmissions crosstalk CO RT User 1: 3km 4km 5km

25 Convergence Guarantee Theorem: ASB algorithm (under high SNR approximation, which leads to frequency-dependent waterfilling) converges to the unique fixed point under both sequential and parallel updates, if the crosstalk channels satisfy (physical meaning also obtained): max n m,k αk n,m < 1 N 1 Recover the convergence of iterative water-filling as a special case Convergence independent of reference line parameters Performance robust to reference line parameters Extensions: ASB for asynchronous transmissions with inter-carrier-interference

26 Amplitude

27 Multiplexing and Shaping Question: How aggressive can we exploit burstiness of triple play (voice, data, video) traffic? Objective: squeeze maximum number of flows into the network, subject to the statistical QoS requirements Previous work in wireline network focus on fixed link rates DSL network has the capability to shuffle the underlying link rates How does this impact the statistical multiplexing decisions? How to do admission control?

28 Example of Problem Formulation Transform stochastic traffic into Effective Bandwidth (EB) The value of EB depends on the traffic characteristics, buffer allocation, and QoS requirement maximize X i w i a i n i (total weighted throughput) subject to n i ν i (ɛ i, B i ) c i, i X B i = B, c C variables n, c, B 0 i (EB less than allocated rate) (total buffer constraint) (capacity region constraint) (# of flows, rate, buffer)

29 Example of Algorithm Two-stage Alternate Maximization (AM) Algorithm: Rate Allocation stage (for fixed buffer B): reduce to weighted rate maximization, can be solved by ASB (autonomous and low complexity) Buffer Allocation stage (for fixed rate c): reduce to quasi-concave maximization, can be solved by bi-section search (in general needs centralized coordination) Alternate through two stages until no further improvement can be obtained Theorem: AM algorithm converges

30 Numerical Example User 1: CO 5km crosstalk Total Number of Flows User 2: RT 4km 3km downstream transmissions ASB (statistical) ASB (deterministic) Both users packet loss probability

31 Space

32 Overview Fat-tree access network topology: Make it robust and survivable Make it economically viable Question: How to add a few links to the tree to make it survivable and economically viable? Three major components of design: Graph theory problem: determine survivable topology (this talk) Optimization problem: allocate bandwidth (another talk) CS systems problem: design real-time signalling protocol (Fraser Lab)

33 Fat Tree Topology Survivable access network design different from backbone network

34 Variations Along Four Dimensions Fat-tree exists or not Single level or multi-level tree Optimization (objective-constraints) model: Minimize total cost with connectivity requirement (e.g. r i edge-disjoint paths from remote terminal i to root) Maximize survivability-based revenue (eg, proportional to number of backup paths) with limited budget (r i is variable) Link cost model: Concave edge cost model: buy-at-bulk Uncapacitated fixed cost: dominant construction

35 Cost Models As a function of distance: affine or convex As a function of link capacity: concave or constant (a) General Concave Cost Model (b) Uncapacitated Fixed Cost Cost Cost Capacity Capacity

36 A World of Graph Theory Problems A taxonomy of 16 problems 4 dimensions of variations, 2 possibilities each Some are difficult (NP-hard) and some are under-explored Two case studies shown here

37 Budget-Constrained Revenue Maximization Uncapacitated fixed cost model, no existing tree, multiple levels NP-hard maximize subject to X h v r v v S X (v,i) b E X f v v,i r v + 1 f v i,j = X (i,j) E b (j,k) E b x e f v i,j f v j,k (total weighted survivability) (number of disjoint paths) (intermediate flow conservation) (e: undirected edge of (i, j)) B X e E c e x e (budget constraint) variables r v, f v i,j 0, x e {0, 1} (survivability, flow, edge selection)

38 Numerical Example Min cost, No survivability (Cost=26, Survivability=0) Min cost, Full survivability (Cost=38, Survivability=8) 8 Budget = 26 8 Budget =

39 Numerical Example Max-revenue access network with limited budget Partial survivability (Cost=33, Survivability=6) 8 Budget =

40 Provisioning Survivability for Existing Single-level Tree Min-cost incremental topology design to provide full survivability Uncapacitated fixed cost model, tree exists, single level Equivalent to Terminal Backup problem: given (required) terminals, Steiner (optional) vertices, and weighted edges, find the cheapest subgraph where every terminal is connected to at least one other terminal (for backup purpose) Polynomial time

41 Numerical Example Established tree in dots, optimized addition of backup links in solid 8 Budget =

42 Time

43 Taxonomy of Problems Question: Which point on rate region boundary to strike at? QoS requirement Characteristic Control mechanism Average throughput Statistical Multiuser scheduler Average delay Statistical Multiuser scheduler Hard delay bound Deterministic Priority queueing Packet loss Statistical Priority queueing & Adm. Ctrl. Inter-user fairness Deterministic Multiuser scheduler & Adm. Ctrl. Multiuser scheduling provides guarantees at the inter-user level Priority queueing provides guarantees at the intra-user level

44 Multiuser Revenue Based Scheduling Multiuser scheduling in MIMO channel with different QoS characteristics Binder Cable Voice User 1 TX RX Video Crosstalk TX RX Data User 2 Multiuser scheduling MIMO channel

45 Example of Problem Formulation maximize R (Flow base rate) subject to β k R R k (s)x k, k = 1,..., K, (Rate constraint) KX x k 1, (Time share constraint) k=1 NX n=1 s k (n) P k, k = 1,..., K, (Power constraint) x k 0, k = 1,..., K, s k (n) 0, k = 1,..., K, n = 1,..., N, variables: R, x k, s k, k = 1,..., K. (1)

46 Joint Time Frequency Scheduling Algorithm Multiuser scheduling with central coordination Channel information Dynamic Spectrum Management v Multiuser Scheduler Flow information R w Priority queueing Transmission of flows in time subproblem Dynamic spectrum management in frequency subproblem

47 Using Processor Sharing Model The PS model serves as a theoretical benchmark for stochastic performance metrics such as average delay A larger revenue corresponds to a larger flow throughput for each user User 1 User 2 Processor sharing Priority queueing differentiates application traffic flows in each user

48 Summary

49 Conclusion and Future Work All three things at the same time: Presents intellectually challenging research issues in broadband access networking Motivates many new and difficult problems in optimization theory, information theory, signal processing, networking, graph theory, stochastic systems Offers an opportunity to make visible, tangible impacts to practical deployment Next step: Empirical data verification of network algorithms Next step: More solutions to this array of research problems

50 The Promise of FAST Copper Broadband Access Rate: Fast Reach: Ubiquitous Reliability: Survivable Quality: QoS for triple play

51 Contacts chiangm